Deep Learning Accurately Detects Atherosclerotic Carotid Vessels in Complex 3D-MERGE Scans: Study

While cardiovascular diseases (CVDs), specifically atherosclerosis-induced ischemic strokes, remain a leading cause of global mortality, previous automated diagnostic research has predominantly focused on traditional black-blood magnetic resonance imaging, leaving a critical clinical gap in the exploration of complex, monochromatic 3D-MERGE imaging modalities. To address this diagnostic limitation, researchers R. Amrit and Anu Shaju Areeckal from the Manipal Academy of Higher Education aimed to thoroughly investigate deep learning-based models, comprehensively evaluating different preprocessing techniques for the precise automated segmentation of carotid artery vessel walls.

Therefore, the researchers evaluated the segmentation accuracy of three deep learning models (U-Net, Attention U-Net, and Residual U-Net) using 5,768 clear images from 1,000 recent stroke patients. The models' performance was assessed using key metrics such as the Dice score, Jaccard index, sensitivity, and specificity.

Key Clinical Findings of the Analysis Includes:

• Superior U-Net Performance: The research demonstrated that the standard U-Net model significantly outperformed others in initial testing, yielding the highest comparative Dice score of 70.85% versus 67.04% for the Attention U-Net without requiring any initial image preprocessing.

• Impact of Preprocessing: The assessment revealed that actively applying Gaussian filtering combined with Contrast Limited Adaptive Histogram Equalization (CLAHE) dramatically enhanced imaging clarity, reducing erroneous segmentation detections by approximately 52% in both the U-Net and Attention U-Net diagnostic frameworks.

• Attention U-Net Enhancement: The analysis indicated that utilizing the specific CLAHE enhancement technique markedly improved the Attention U-Net model's anatomical precision, successfully increasing its predictive Dice score from an initial 67.04% to an optimized 70.51%.

• Cross-Validation Robustness: The investigation highlighted that during a rigorous fivefold cross-validation study, the dynamic Attention U-Net model exhibited greater overall resilience to diverse data variations, marginally outperforming the standard U-Net with an average Dice score of 80.01% versus 79.77%.

• Lumen Area Estimation: The findings confirmed that the efficiently preprocessed U-Net model achieved the most accurate direct measurement of the inner vessel opening, displaying a minimal comparative error deviation of only 2.39 square millimeters versus 3.36 square millimeters for the Attention U-Net.

The results suggest that standard U-Net and Attention U-Net models hold tremendous clinical potential for the robust automated segmentation of carotid vessel walls in monochromatic 3D-MERGE images, achieving Dice scores up to 80% when actively augmented by advanced image preprocessing techniques.

Thus, the analysis concludes that by seamlessly integrating these sophisticated image processing architectures, clinicians may soon have access to highly accurate automated tools for precisely measuring vessel lumen areas, gently facilitating the earlier evaluation of potentially fatal atherosclerotic plaques with minimal manual diagnostic intervention.

Although the current iteration of these models lacks the complete ability to automatically localize the initial region of interest and processes external and internal arteries indiscriminately, systematically incorporating additional automated preprocessing methods alongside larger, diverse clinical datasets could seamlessly refine these deep learning architectures for an even broader future diagnostic applicability.

Reference

Amrit, R., & Areeckal, A. S. (2026). Performance Analysis of Deep Learning Models for Segmentation of Carotid Artery Vessel Wall in 3D-MERGE Images. Indian Journal of Radiology and Imaging, 36(3), 394–404.